The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local scripts to perform its core functions. It calls python scripts/search_docs.py to index and search documentation based on user queries and provides a mechanism to run node ./bin/linter-cli.js for code diagnostics. These commands are restricted to the local environment provided within the skill package.
[EXTERNAL_DOWNLOADS]: The linter-cli setup instructions include an installation step (npm run install-runtime-deps) that performs a shallow clone of a repository from https://gitcode.com/openharmony/third_party_typescript.git. This download is performed to fetch a customized TypeScript runtime required for ArkTS linting. GitCode is a well-known service in the context of OpenHarmony development.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its documentation search feature.
Ingestion points: Untrusted user input enters the agent context via the --query argument passed to scripts/search_docs.py in SKILL.md.
Boundary markers: The skill instructions do not specify any delimiters or special markers to wrap search results or user queries to prevent the agent from obeying instructions embedded in the data.
Capability inventory: The skill has the capability to execute shell commands (Python and Node.js), write to the local file system (via the linter cache), and perform network operations (via npm and git during setup).
Sanitization: There is no evidence of sanitization or filtering logic in the skill instructions to process external content before it is presented to the agent.

hmos-arkts-knowledge-retriever